FirstRead / From answers to a category
From Jev’s questions to risk levels
A yes/no question gets a probability of “yes”, not just a binary answer. FirstRead’s rules turn those scores into a level.
Recorded message: “i've decided, it's tonight. i've already written the letters”
1 Ask a yes/no question
“Do they intend to act on those thoughts?”
2 Jev returns P(yes)
0.96≈ 96% “yes”
≈ 4% “no”
3 Compare with a policy cutoff
0.96 ≥ 0.35 → CRISIS
This intent rule matches.
| Question Jev answers | P(yes) | WATCH | ELEVATED | CRISIS |
|---|---|---|---|---|
| Do they express thoughts of dying or hurting themselves? | Jev’s P(yes)0.94 | WATCH✓ Met: ≥ 0.15 | ELEVATED✓ Met: ≥ 0.40 | CRISIS— |
| Are they harming themselves, or feeling urges to do so? | Jev’s P(yes)0.85 | WATCH✓ Met: ≥ 0.15 | ELEVATED✓ Met: ≥ 0.30 | CRISIS— |
| Do they intend to act on those thoughts? | Jev’s P(yes)0.96 | WATCH— | ELEVATED✓ Met: ≥ 0.20 | CRISIS✓ Met: ≥ 0.35 |
| Have they made preparations, such as goodbye letters? | Jev’s P(yes)0.96 | WATCH— | ELEVATED— | CRISIS✓ Met: ≥ 0.40 |
| Might they act soon, such as today or tonight? | Jev’s P(yes)0.97 | WATCH— | ELEVATED✓ Met: ≥ 0.20 | CRISIS✓ Met: ≥ 0.40 |
NONE if no policy rule matches.
Otherwise, take the highest match: WATCH < ELEVATED < CRISIS.
This turn: CRISIS → escalateStop generation, send the crisis reply, emit a handoff event.
P(yes) is Jev’s answer probability, not the probability of harm. Five user questions shown; the full policy includes combined rules, and conversation history can retain a higher level. All 18 questions (14 user, 4 reply) (source on GitHub, opens in a new tab) · Full policy walkthrough.